"Alexa, turn off the lights." It should be just as easy with ioBroker: You say "ioBroker," and your smart home listens. And it does so completely locally, on a Raspberry Pi or a small ESP32 speaker, without your voice ever leaving the house.
We need your help for this. More precisely: your vote.
👉 Join now at wake.iobroker.in
What is it about?
In our article about AI in ioBroker, we introduced the Smart Assistant, which allows you to control your smart home via voice commands. Before the assistant listens, however, the voice device must first recognize that it is being addressed. This is handled by the activation word, also known as the "wake word."
This recognition process runs constantly in the background. Therefore, it belongs on the device itself: only when the word has been recognized is anything recorded and processed. Everything before that remains unused.
For the models, we are relying on two independent projects:
- openWakeWord for the Raspberry Pi, for example with the satellite
@iobroker/assistant-satellite, - microWakeWord for small ESP32 speech devices that work with ESPHome.
Why real voices?
An activation word must work for everyone: with deep and high voices, with fast and slow speech, with dialect or accent, on a headset as well as on a mobile phone from a few meters away. At the same time, it must not be triggered by every similar-sounding word.
Training data can also be artificially generated. However, such voices are almost exclusively available in English, and they all sound quite similar. A German-pronounced "ioBroker" in all its variations is impossible to achieve this way. For that, you need real voices, and preferably many different ones.
This is how it works
Participation takes less than a minute and can be done directly in your browser, on your PC, laptop or smartphone. No installation is required.
- Specify the pronunciation. Write down how you pronounce "ioBroker": "IO-Broker", "Ei-O-Broker", in one go... If you say it one way sometimes and another way other times, you can specify up to three pronunciations.
- Record it. Then record the word a few times, three seconds each time: at normal speed, quickly, slowly, quietly, and from a distance. You can stop after three recordings, but every additional one helps.
- Similar words (optional). If desired, you can also record a few similar-sounding words. This will teach the model not to react to "Broker" or "Joker".
A few tips for good recordings: Find a room that is as quiet as possible, speak only after pressing the button, and listen briefly before saving.
Especially sought after
The more diverse the voices, the better the model. We welcome women and men, young and old, dialects and accents, voices from Germany, Austria, and Switzerland. Feel free to use a variety of microphones as well: headset, laptop microphone, or mobile phone.
The whole family is welcome to participate. There's an option at the end called "Someone else will join in." Each person must give their own consent. You can participate if you are 16 or older.
Data protection
It's about your voice, which is why this part is particularly important to us:
- No names, no email addresses. We don't ask for either. Information such as age, gender, native language, region, and microphone is voluntary and can be left set to "no comment." However, it helps us to keep track of the diversity of voices.
- The recordings remain internal. They will not be published and are used only for training and testing the models. Only the finished models will be published, and recordings cannot be recovered from them.
- Data is stored in the EU. It resides in a non-public storage facility at AWS in Ireland.
- You retain control. After giving your consent, you will receive a deletion code that allows you to delete your recordings yourself at any time. Please keep it safe: without the code, we cannot identify your recordings. Models that have already been trained will not be affected by the deletion.
What happens next?
Once enough data has been collected, we will train the models. We will make the finished models available in the forum and directly in the adapters. Then ioBroker will respond to its own name.
Questions, problems or ideas? Feel free to write to us in the thread in the forum .
Thanks for participating! 🎙️